Limitations of and extensions to heuristic search planning

• Main Author: Burfoot, Daniel.
• Format: Others
• Language: en
• Published: McGill University 2006
• Subjects: Heuristic programming.
• Online Access: http://digitool.Library.McGill.CA:80/R/?func=dbin-jump-full&object_id=100779

This thesis explores limitations of heuristic search planning, and presents techniques to overcome those limitations. The two halves of the thesis discuss problems in standard propositional planning (STRIPS) and in planning with numeric state variables respectively. === In the context of STRIPS, the primary focus is on the widely used relaxed plan heuristic (h+). A variety of cases are shown in which h+ provides systematically bad estimates of goal distance. To address this breakdown, a planning system called RRT-Plan is presented. This system is inspired by the concept of Rapidly-exploring Random Trees, which was originally developed for use in mobile robot path planning. Experimental results show that RRT-Plan is comparable to leading planners in terms of number of problems solved and plan quality. We conclude that the effectiveness of RRT-Plan is based on its ability to search the space of artificial goal orderings. === The second half of the work considers heuristic search planning in numeric domains. Two particularly significant obstacles are identified. The Curse of Affluence is due to the vast blowup in the search space caused by the addition of numeric variables. The Curse of Poverty relates to the difficulty of finding relevant lower bounds on resource consumption. === Exploration of the Curse of Affluence leads to the new concepts of reduced search and enhanced states. In reduced search, certain simple operators are not used to expand states. Instead, enhanced states are constructed which represent all possible states which could be achieved by suitably inserting simple operators in the plan. Enhanced states are represented by a set of constant discrete variables, and a convex hull of numeric values. This representation can be queried and updated in a natural way. Experimental results show that there are domains for which reduced search gives order of magnitude performance improvements over Metric-FF, a leading heuristic search planner for numeric domains.